Unsupervised AI-based water consumption classification from time series data

Abstract Increasing concern about global water scarcity has highlighted the necessity to optimize the use of this resource, where AI and machine learning techniques can play a key role, serving as decision-making support tools that would help anticipate necessities and ensure a quality supply. This study proposes a methodology to characterize residential water users from daily consumption time series by applying feature extraction, dimensionality reduction and clustering techniques. Aggregated statistical features were first computed and then reduced via PCA. DBSCAN outperformed K-means in clustering, and both UMAP and t-SNE showed adequate embedded space results. Different characteristics were found between clusters, demonstrating the possibility of separating users of the same grid by behaviors or profiles.